Prediction of supercritical CO2 heat transfer behaviors by combining transfer learning and deep learning based on multi-fidelity data

The flow and heat transfer characteristics of supercritical CO₂ are important for heat exchanger design and the safe operation of supercritical CO₂ power cycles. However, it is difficult to predict the supercritical heat transfer behaviors due to the non-monotonic temperature distribution in the case of the heat transfer deterioration (HTD) phenomenon. For low-cost, fast and accurate prediction of the supercritical heat transfer behavior, this study proposed a transfer learning model based on multi-fidelity data to achieve fast prediction with acceptable accuracy over a wide range of working conditions. This method fully utilized the low fidelity data (empirical correlations) and the medium fidelity data (numerical results) to generate a large amount of data for pretraining, in which the Latin Hypercube Sampling (LHS) method combined with the HTD correlation was used for sampling. For fine-tuning, high fidelity data from experiments was employed. Compared to the deep learning model trained directly with high fidelity dataset, the transfer learning model demonstrated vastly improved predictive performance on both the test and validation datasets. Additionally, the coefficient of determination R² was discussed to preventing from “physical overfitting”. Instead of excessively pursuing the high R² (close to 1), the validity of the prediction should be concerned, especially when using the non-smooth experimental data as the dataset for model training. Moreover, the trained models and the relative files are available at Supplementary materials.